NaN passed to gcvt

[p5pRT]

Migrated from rt.perl.org#3100 (status was 'resolved')

Searchable as RT3100$

[p5pRT]

From gin@mo.msk.ru

Created by gin@mo.msk.ru

lib/bigfltpm test fails in step 351. `perl' passes a NaN value to gcvt()\, which causes it to get SIGFPE and dump core. gdb backtrace follows.

#0 0x80019223 in PL_regflags () #1 0x80f721d in cvt () #2 0x80f71ae in ecvt () #3 0x80f7025 in gcvt () #4 0x80aa5bb in Perl_sv_2pv (sv=0x811cb50\, lp=0x804787c) at sv.c​:2114 #5 0x80a23ce in Perl_pp_match () at pp_hot.c​:956 #6 0x809f74c in Perl_runops_standard () at run.c​:27 #7 0x805bd66 in S_run_body (oldscope=1) at perl.c​:1401 #8 0x805ba3b in perl_run (my_perl=0x8116804) at perl.c​:1325 #9 0x80590ee in main (argc=2\, argv=0x8047b18\, env=0x8047b24) at perlmain.c​:52 #10 0x8058feb in .text ()

Hex dump of double value in question is​:

0x00 0x00 0x00 0x00 0x00 0x00 0xf8 0x7f

It always causes my gcvt to get SIGFPE\, regardless of program which passed such an arg to it.

I do not know\, is such a behavior of perl considered proper (and gcvt must handle NaN's) or it is a bug in perl.

Perl Info

``` Flags: category=core severity=low Site configuration information for perl v5.6.0: Configured by gin at Fri Apr 14 15:07:46 MSD 2000. Summary of my perl5 (revision 5.0 version 6 subversion 0) configuration: Platform: osname=sco, osvers=3.2v5.0.2, archname=i386-sco uname='sco_sv d-fens 3.2 2 i386 ' config_args='-de' hint=recommended, useposix=true, d_sigaction=define usethreads=undef use5005threads=undef useithreads=undef usemultiplicity=undef useperlio=undef d_sfio=undef uselargefiles=define use64bitint=undef use64bitall=undef uselongdouble=undef usesocks=undef Compiler: cc='cc', optimize='-O0', gccversion= cppflags='-U M_XENIX -D PERL_SCO -D PERL_SCO5 -w0 -belf -I/usr/local/include' ccflags ='-U M_XENIX -D PERL_SCO -D PERL_SCO5 -w0 -belf -g -I/usr/local/include' stdchar='unsigned char', d_stdstdio=undef, usevfork=false intsize=4, longsize=4, ptrsize=4, doublesize=8 d_longlong=undef, longlongsize=, d_longdbl=define, longdblsize=12 ivtype='long', ivsize=4, nvtype='double', nvsize=8, Off_t='off_t', lseeksize=4 alignbytes=4, usemymalloc=y, prototype=define Linker and Libraries: ld='ld', ldflags =' -L/usr/local/lib' libpth=/usr/local/lib /lib /usr/lib /usr/ccs/lib libs=-lintl -lsocket -lnsl -lndbm -ldbm -lld -lm -lc -lcrypt -lsec -lPW -lx libc=, so=so, useshrplib=false, libperl=libperl.a Dynamic Linking: dlsrc=dl_dlopen.xs, dlext=so, d_dlsymun=undef, ccdlflags='-Wl,-Bexport -L/usr/local/lib' cccdlflags='-Kpic', lddlflags='-G -L/usr/local/lib' Locally applied patches: @INC for perl v5.6.0: lib /usr/local/lib/perl5/5.6.0/i386-sco /usr/local/lib/perl5/5.6.0 /usr/local/lib/perl5/site_perl/5.6.0/i386-sco /usr/local/lib/perl5/site_perl/5.6.0 /usr/local/lib/perl5/site_perl . Environment for perl v5.6.0: HOME=/home/gin LANG=C_C.C LANGUAGE (unset) LD_LIBRARY_PATH (unset) LOGDIR (unset) PATH=.:/home/gin/pbin:/bin:/usr/bin:/usr/local/bin:/usr/merge/dosroot/ubin:/usr/bin/X11:/home/gin/bin PERL_BADLANG (unset) SHELL=/bin/sh ```

[p5pRT]

From [Unknown Contact. See original ticket]

One possible solution of gcvt SIGFPE problem is using sprintf instead (in Gconvert macro). It runs normally and returns "nan". After fixing config.sh this way and rebuilding perl all tests succeed.

This problem is specific to 5.0.2 or earlier versions. Perhaps this explains why it was not detected before​: perl is likely to have been ported only on 5.0.4 or later (-Bexport cc flag is valid only there).

[p5pRT]

From @gsar

On Fri\, 14 Apr 2000 17​:32​:06\, gin@​mo.msk.ru wrote​:

It always causes my gcvt to get SIGFPE\, regardless of program which passed such an arg to it.

I do not know\, is such a behavior of perl considered proper (and gcvt must handle NaN's) or it is a bug in perl.

Your platform likely needs to disable SIGFPE. See the end of unixish.h for examples.

Sarathy gsar@​activestate.com

[p5pRT]

From @schwern

Is this still an issue with 5.8.6?

[p5pRT]

@schwern - Status changed from 'open' to 'stalled'

[p5pRT]

From gin@mo.msk.ru

Again\, already posted the solution (\023a38fa0f7b94\-gin@​mo\.msk\.ru). `Gconvert' must be implemented on version 5.0.2 or earlier using `sprintf'\, not `gcvt'. Again\, check `hints/sco5.sh' or whatever sets default values for `Configure' on that system.

Moreover\, one can reproduce the way `Gconvert' is called by batcktrace from original bug report (\023a38f7561586\-gin@​mo\.msk\.ru) and write a test to be used by `Configure' / `hints' script to check how to implement `Gconvert' properly.

[p5pRT]

The RT System itself - Status changed from 'stalled' to 'open'

[p5pRT]

From @doughera88

On Tue\, 14 Dec 2004\, Ilya N. Golubev wrote​:

Again\, already posted the solution (\023a38fa0f7b94\-gin@​mo\.msk\.ru). `Gconvert' must be implemented on version 5.0.2 or earlier using `sprintf'\, not `gcvt'. Again\, check `hints/sco5.sh' or whatever sets default values for `Configure' on that system.

Yes\, that's in the bug database.

However\, the solution Sarathy suggested (ignoring SIGFPE) is also probably already implemented in 5.8.6 (see the end of perl.h)\, though I don't know if that implementation is working on SCO 5.0.2\, and I expect nobody is in a position to test it.

Moreover\, one can reproduce the way `Gconvert' is called by batcktrace from original bug report (\023a38f7561586\-gin@​mo\.msk\.ru) and write a test to be used by `Configure' / `hints' script to check how to implement `Gconvert' properly.

Configure does indeed contain a framework for testing various inputs to the Gconvert function and verifying that they work\, but that framework doesn't handle NaN values\, so it's not trivial to simply add in your suggested test. It's a good suggestion\, but would take a bit of work.

Still\, the following patch should be harmless and safe to apply. It simply instructs Configure to prefer sprintf() on SCO.

Inline Patch

```diff diff -r -u perl-current/hints/sco.sh perl-andy/hints/sco.sh --- perl-current/hints/sco.sh Sun Jun 24 10:21:37 2001 +++ perl-andy/hints/sco.sh Wed Dec 15 10:13:25 2004 @@ -201,6 +201,15 @@ fi ############################################################### +# At least for ORS5.0.2, prefer sprintf() over gcvt(), since gcvt() +# used to cause a SIGFPE and a core dump when passed a NaN. +# This may not be an issue in perl-5.8.x and later since we +# try to trap SIGFPE. However, preferring sprintf() should be +# safe anyway, so let's go ahead and set it. See the bugs database +# item [perl #3100]. --A.D. 12/2004. + gconvert_preference='sprintf' + +############################################################### # We disable ODBM_File if OSR5 because it's mostly broken # but keep it for ODT3 as it seems to work. if test "$scorls" = "5"; then -- ```

  Andy Dougherty doughera@​lafayette.edu

[p5pRT]

From @rgs

Andy Dougherty wrote​:

Still\, the following patch should be harmless and safe to apply. It simply instructs Configure to prefer sprintf() on SCO.

diff -r -u perl-current/hints/sco.sh perl-andy/hints/sco.sh --- perl-current/hints/sco.sh Sun Jun 24 10​:21​:37 2001 +++ perl-andy/hints/sco.sh Wed Dec 15 10​:13​:25 2004

Thanks\, applied as #23655.

[p5pRT]

@rgs - Status changed from 'open' to 'resolved'